NumPy contains core numerical functionality. The main thing it provides is the ndarray object. It also provides things like a high quality pseudorandom number generator (an implementation of the Mersenne twister, with generators for several distributions), fast Fourier transforms, scalar types that correspond to underlying C data types and make their size explicit (e.g. float64, int32, complex128), a variety of ufuncs for things like trigonometric functions and transcendentals, linear algebra routines (that can use an optimized BLAS such as ATLAS or the Intel Math Kernel Library if you have them available), and lots of utility functions for working with ndarrays: creating them, merging them, splitting them, etc. It also provides f2py which is a tool for generating Python wrappers for Fortran code.

SciPy is a much broader package that is best thought of as a set of toolboxes, built on top of NumPy, and also wrapping suitably licensed packages (including many numerical libraries written in Fortran using f2py). Things like clustering, signal processing, even more linear algebra routines than NumPy provides, sparse matrix support, numerical optimization; all of this stuff falls under SciPy. You're best looking through the SciPy Reference Guide for more detailed guide to what is included.

One notable difference is that NumPy has as part of its requirements that it can be built without a Fortran compiler; SciPy depends heavily on Fortran and you'll need a Fortran compiler such as gfortran, g77 or ifort to compile it.